Polymorphic Case study
← All case studies

01 / INSURANCE / ILLUSTRATIVE

Six hours of medical review, compressed to 45 seconds.

A privacy-preserving extraction and validation pipeline that transformed unstructured medical records into auditable underwriting evidence—without turning the model into the decision-maker.

The plateRecord: a medical file read page by page. Facts: each extracted with the line it came from, then challenged by an independent pass. Escalation: a fact the challenger cannot confirm goes to the underwriter with both readings and their sources. The record and its facts are invented.

ClientANONYMISED MUTUAL INSURER
CapabilityDOCUMENT INTELLIGENCE
Duration18 WEEKS
StatusILLUSTRATIVE CASE

01 / the decision
What information does an underwriter need—and how certain are we?

The bottleneck was reliable interpretation.

Experienced underwriters were spending up to six hours navigating long, inconsistent medical files. Important diagnoses, medications, dates, and risk indicators could appear in scanned letters, tables, abbreviations, or conflicting clinical notes.

The organisation wanted faster turnaround, but the workflow was too consequential for an opaque summary. Every extracted fact needed provenance, confidence, and a clear route back to the source.

Fig. 01

Processing time per file

From six hours of manual review to under a minute of preparation before the underwriter sees the case.

Lower is better

02 / the build
Ground → extract → challenge → review

A pipeline with a built-in second opinion.

We combined layout-aware parsing with a constrained extraction layer that mapped diagnoses, medications, procedures, dates, and risk flags into a typed schema. Every field retained page-level provenance.

A separate challenger model tested the evidence for contradictions, temporal ambiguity, and unsupported inference. Low-confidence or disputed fields were routed to an underwriter with the relevant source passages already assembled.

  1. Stage 0101

    De-identify

    Replace direct identifiers with reversible encrypted tokens before model processing.

  2. Stage 0202

    Structure

    Parse scans, tables, clinical notes, and attachments into a source-addressable record.

  3. Stage 0303

    Extract & challenge

    Generate candidate facts, then independently test support, chronology, and contradictions.

  4. Stage 0404

    Escalate

    Send uncertainty—not volume—to the human reviewer with evidence already attached.

Fig. 02

Agreement with expert review, over eight eval rounds

Agreement climbed from 78.2% to 95.3%, crossing the 94% bar the pilot was held to in round seven.

Target ≥ 94%

03 / the result
Measured in time, agreement, and privacy

Faster where it was safe. Human where judgement mattered.

The production pattern separated evidence assembly from underwriting judgement. The system handled document navigation and consistency checks; authorised professionals retained the decision and could inspect every claim.

480×

faster evidence preparation

95.3%

agreement with expert review

0

reported PHI exposure incidents in the pilot

The same architecture now serves as a reusable pattern for other high-volume, high-sensitivity records where claims must remain traceable to source.